Problem

Sciloop addresses a bottleneck in frontier-AI development: models are becoming strong enough to perform well on conventional tests, while labs still need genuinely difficult, expert-authored problems that expose failures in mathematical and scientific reasoning. The company says its problems are created by the top 0.01% of STEM talent, and Y Combinator reports that GPT 5.4 Pro and Gemini 3.1 Pro score only 0–5% on Sciloop’s hardest problems. The killer use case is therefore supplying frontier AI labs with fresh math and physics problems, solutions, and benchmarks for model training and evaluation.

The pain is both technical and economic. Labs need credible, continuously refreshed tests and training examples, but expert reasoning is scarce and automated metrics or model self-evaluation can miss important weaknesses. Sciloop’s model makes that scarce labor explicit: experts are paid roughly $100–$300 per data point, or $225 or more for some problems, while labs specify the domain, difficulty, format, and volume they need. For buyers, the intended payoff is more targeted diagnosis of model weaknesses and better data for improving or validating frontier systems.

Product / Service

Sciloop is a managed expert-data service with a two-sided contributor platform. AI labs provide requirements, after which Sciloop handles expert sourcing, quality control, and continuous delivery. On the supply side, qualified contributors create a problem and its expected correct response; Sciloop’s AI pipeline grades submissions for difficulty and expert reviewers vet them before payment. This produces tailored problem-and-solution datasets rather than generic crowdsourced labels.

The delivery model is designed to support both training and evaluation. Sciloop presents active expert problem sets, verified solutions, and continuously updated data streams, allowing a lab to revise its requirements as it learns more about a model’s capabilities. The benefit is access to difficult, domain-specific reasoning data without a lab having to recruit and manage every Olympiad medalist, researcher, or specialist itself.

Market

Sciloop competes in the overlapping markets for expert human data, AI model training data, frontier-model evaluation, and reasoning benchmarks. Y Combinator categorizes it across artificial intelligence, marketplaces, data labeling, and big data. Its closest established alternatives are broader providers rather than exact replicas: Scale’s Scale Evaluation offers model comparison, error discovery, and improvement guidance based partly on human evaluations, while Surge AI positions itself around supplying human intelligence for AGI. Sciloop’s claimed differentiation is the narrow focus on exceptionally difficult, expert-created STEM reasoning problems and continuously refreshed bespoke data.

The company appears to be an early commercial-stage startup, not a demonstrably revenue-scale business. Y Combinator lists it as founded in 2025, part of Fall 2025, active, and staffed by two founders; its profile also says it has been used in MIT research groups and by early partners. The website shows active project and review activity, including a displayed stream with 12 items in review, but another public platform page says the platform is still coming soon. No revenue figures or named AI-lab customers are disclosed in the retrieved company materials. Third-party databases report $500,000 raised from E14 Fund and Y Combinator, which is evidence of financing but not proof of revenue or product-market scale.